Stimulus-locked labels detect movement above chance but underperform onset-locked labels in seven SEEG patients

In an offline analysis of seven SEEG patients performing random-dot motion decisions with mouse movement, stimulus-locked calibration labels detected movement above chance but did not meet a 0.05-AUC non-inferiority margin against onset-locked labels. With Ledoit-Wolf shrinkage LDA on 70–150 Hz features, mean AUC was 0.694 for stimulus-locked labels and 0.788 for onset-locked labels (mean difference −0.094; three of seven patients within the margin; p = 0.76, Holm-adjusted p = 0.76). Stimulus-locked labels from training trials with reaction time at least 0.5 s, so the label window contained no movement onset, still detected movement in all seven patients (mean AUC 0.685; p = 0.0078, Holm-adjusted p = 0.016). This detector was 0.095 AUC lower than an onset-locked detector trained on the same trials (all seven patients lower; two-sided p = 0.031). Across 0.2 s bins, the decoders were similar from 0.4 s before movement onset to onset and lower for stimulus-locked labels from onset through 1 s after. These findings suggest that stimulus-locked labels mainly train on decision and preparation activity rather than movement execution. The result is limited to offline overt mouse movement, seven patients, one session and time-ordered split per patient, one feature and one decoder; it does not establish performance for attempted movement by paralyzed users.

brain-computer interface · SEEG · movement detection · calibration labels · stimulus-locked · movement onset · shrinkage LDA · high-gamma activity · AUC · non-inferiority